Traditional machine learning treats brain data as if it were a flat photograph or a simple spreadsheet. When we apply a standard Convolutional Neural Network (CNN) to an EEG or fMRI scan, the model assumes that the "pixels" (sensors or voxels) are arranged on a uniform grid where the only thing that matters is physical proximity.
However, as Yufei Huang points out in Deep Learning Models for EEG-based Rapid Serial Visual Presentation Event Classification, this "grid-like" assumption is a fundamental mismatch for neurobiology. The human brain is a highly folded, non-Euclidean manifold. Two regions might be physically close in 3D space but belong to entirely different functional networks, separated by a deep cortical fold. Conversely, two distant regions might be "wired" together by long-range white matter tracts, functioning as a single unit.
To capture how the brain actually processes information, researchers are moving away from Euclidean grids and toward Geometric Deep Learning, treating the brain as a structured, dynamic graph.
The Geometry of the Scalp and Cortex
The first hurdle in brain analysis is the physical container. EEG sensors are placed on a spherical scalp, and the underlying cortex is a complex, crumpled surface. Pengcheng Xue and colleagues in CortiLife: A Unified Framework for Cortical Representation Learning across the Lifespan argue that standard 3D convolutions fail because they cannot respect this intrinsic topology.
When we flatten these 3D relationships into 2D images, we lose critical structural information. Haoning Wang addresses this in KAST-BAR by using a "3D-aware brain topology encoder." Similarly, Matthew Lyon and Paul Armitage developed the Parametric Continuous Convolutional Neural Network (PCCNN), which uses continuous convolutions to natively process the spherical manifold of diffusion MRI data. By respecting the actual geometry of the head, these models achieve higher accuracy with significantly fewer parameters.
From Sensors to Graphs
If the brain isn't a grid, what is it? The consensus across recent literature is to treat it as a graph. In this view, each sensor or brain region is a "node," and the relationships between them are "edges."
Graph Neural Networks (GNNs) allow models to aggregate information along these edges, regardless of how far apart the nodes are in Euclidean space. Jingcong Li demonstrates this with the Functional Graph Discriminative Network (FGDN), which models brain connectivity as a structured graph to identify biomarkers for Autism Spectrum Disorder.
The intuition here is simple: if the frontal lobe and the parietal lobe are communicating during a task, a GNN can "hop" directly between those nodes. A standard CNN would have to filter through all the irrelevant "empty space" between them.
Learning the Dynamic Wiring
The most powerful versions of these models don't just use a fixed map of the brain; they learn the connections. This is known as adaptive or dynamic graph learning.
In GraphSleepNet, Ziyu Jia and colleagues developed a framework that adaptively learns the "intrinsic connections" between EEG channels. Instead of assuming the connections are based on physical distance, the model looks at which channels are synchronizing in real-time.
This concept is pushed further by Yu-Kai Wang in the Connectivity-Aware Graph Neural Network (CAGNN). By using self-attention mechanisms, the model "wires" the graph specifically for the task at hand—such as detecting drowsiness—identifying which neural pathways are most relevant to the user's state.
Regularizing with Neurobiology
While GNNs provide the flexibility to model non-Euclidean relationships, they can sometimes become too flexible, fitting to noise in the data. To prevent this, researchers use anatomical regularization.
Early work by Rémi Cuingnet in Spatial and anatomical regularization of SVM for brain image analysis introduced Laplacian regularization. This forces the model to respect known anatomical boundaries, ensuring that the resulting classification maps align with neurobiology.
Modern architectures like NeuroGNN by Arash Hajisafi combine this spatial prior with temporal dynamics. By integrating the physical proximity of sensors, the temporal flow of signals, and the semantic function of specific brain regions, these models create a "holistic" representation of neural activity.
Why It Works: The Intuition of Information Flow
The reason geometric learning outperforms traditional methods is that it aligns the model's architecture with the brain's communication protocol.
- Symmetry and Invariance: Just as a CNN is invariant to where an object appears in an image, a GNN can be invariant to the specific "folding" of an individual's brain. Jinpei Han and A. Aldo Faisal showed that by leveraging these geometric symmetries, GNNs can uncover stable "neural signatures" of motor learning that are consistent across different people.
- Capturing Synergies: In wearable applications, such as the finger-movement decoding in SPECTRE, sensors are often arranged in a cylinder around a limb. Zihan Weng argues that modeling this cylindrical relationship allows the network to capture "muscle synergies"—how groups of muscles work together—which a flat sequence model would miss.
- Efficiency: By ignoring the "Euclidean void" and focusing only on functional edges, models like REST by Arshia Afzal can be made incredibly small (37KB) and fast, suitable for real-time brain implants.
By treating the brain as a non-Euclidean manifold rather than a grid, this research direction moves AI closer to the actual "wiring diagram" of human intelligence.